Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “functional modeling”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 613 records · Page 34
All sourcesNASA NTRSDOE OSTI
DOE OSTI2021

Modeling the dynamic response of rock masses with multiple compliant fluid saturated joint sets - Part II: Continuum modeling

Part I of this two-part paper describes enhancements of a mesoscale model to include: many non-persistent fluid saturated joints with finite area, normal compliance with hysteresis and the weakening effects of angular deviation in the joint sets and of fluid pressure in the joints. The objective of this Part II is to use the results of the mesoscale model as simulated data to inspire functional forms in a continuum model for rock masses with multiple compliant fluid saturated joint sets which include these enhancements. Also, an algorithm is proposed for reducing the number of joint sets being modeled. Examples are presented here which show that the continuum model can predict similar anisotropic response to that observed in mesoscale simulations. In contrast with the mesoscale model, the continuum model can be used for large scale simulations of real applications, smoothly transitioning from coupled inelastic thermomechanical response near an explosive source to orthotropic elastic response far from the source.

58 GEOSCIENCES↗
DOE OSTI2024

Active and Transfer Learning of High-Dimensional Neural Network Potentials for Transition Metals

Classical molecular dynamics (MD) simulations represent a very popular and powerful tool for materials modeling and design. The predictive power of MD hinges on the ability of the interatomic potential to capture the underlying physics and chemistry. There have been decades of seminal work on developing interatomic potentials, albeit with a focus predominantly on capturing the properties of bulk materials. Such physics-based models, while extensively deployed for predicting the dynamics and properties of nanoscale systems over the past two decades, tend to perform poorly in predicting nanoscale potential energy surfaces (PESs) when compared to high-fidelity first-principles calculations. These limitations stem from the lack of flexibility in such models, which rely on a predefined functional form. Machine learning (ML) models and approaches have emerged as a viable alternative to capture the diverse size-dependent cluster geometries, nanoscale dynamics, and the complex nanoscale PESs, without sacrificing the bulk properties. Here, in this study, we introduce an ML workflow that combines transfer and active learning strategies to develop high-dimensional neural networks (NNs) for capturing the cluster and bulk properties for several different transition metals with applications in catalysis, microelectronics, and energy storage, to name a few. Our NN first learns the bulk PES from the high-quality physics-based models in literature and subsequently augments this learning via retraining with a higher-fidelity first-principles training data set to concurrently capture both the nanoscale and bulk PES. Our workflow departs from status-quo in its ability to learn from a sparsely sampled data set that nonetheless covers a diverse range of cluster configurations from near-equilibrium to highly nonequilibrium as well as learning strategies that iteratively improve the fingerprinting depending on model fidelity. All the developed models are rigorously tested against an extensive first-principles data set of energies and forces of cluster configurations as well as several properties of bulk configurations for 10 different transition metals. Our approach is material agnostic and provides a methodology to transfer and build upon the learnings from decades of seminal work in molecular simulations on to a new generation of ML-trained potentials to accelerate materials discovery and design.

36 MATERIALS SCIENCE↗
DOE OSTI2024

Self-similar Reynolds-averaged mechanical–scalar turbulence models for reshocked Richtmyer–Meshkov instability-induced mixing in the small Atwood number limit

Analytical self-similar solutions to two-, three-, and four-equation Reynolds-averaged mechanical–scalar turbulence models describing incompressible turbulent Richtmyer–Meshkov instability-induced mixing in planar geometry derived in the small Atwood number limit are extended to construct models for reshocked Richtmyer–Meshkov mixing. In this study, the models are based on the turbulent kinetic energy K and its dissipation rate ε, together with the scalar variance S and its dissipation rate χ modeled either differentially or algebraically. The three- and four-equation models allow for a simultaneous description of mechanical and scalar mixing, i.e., mixing layer growth and molecular mixing. Mixing layer growth parameters and other physical observables were obtained explicitly as functions of the model coefficients and were used to calibrate the model coefficients. Here, the solutions for the singly shocked Richtmyer–Meshkov case for the mixing layer width and the turbulent fields are used to construct piecewise-continuous generalizations of these quantities for times after reshock. For generality, the post-reshock mixing layer width is not assumed to grow with the same power-law as the pre-reshock width, and an impulsive approximation applied to Rayleigh–Taylor instability growth is used to establish the expression for the post-reshock width. A four-equation model is then used to illustrate the spatiotemporal behavior of the mean and turbulent fields and late-time turbulent equation budgets across the mixing layer. The reference solutions derived here can provide systematic calibrations and better understanding of mechanical–scalar turbulence models and their predictions for reshocked Richtmyer–Meshkov instability-induced turbulent mixing in the very large Reynolds number limit.

42 ENGINEERING↗
DOE OSTI2023

pyNuMAD v.0.1

SAND2024-08606O The pyNuMAD software is used for managing wind turbine blade model data. pyNuMAD specializes in defining the geometry, materials, and boundary conditions for structural analysis of wind turbine blades. This includes loading in data files and providing an interface for users to make updates to the model. The software also features meshing functionality, which takes the blade model and creates a shell or brick mesh for use in finite element analysis. A typical user workflow might be: load in blade information from a yaml file, make adjustments to the blade properties, update the blade based on the adjustments, create a mesh of the blade, export this blade to another software for structural analysis. Sandia National Laboratories is a multimission laboratory managed and operated by National Technology & Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell International Inc., for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-NA0003525.

Paquette, Joshua↗
DOE OSTI2024

Machine-learning based model reduction for partial differential equations

We develop a novel synergistic approach between model reduction and machine learning. The specific goal of this project is to aid in the construction of reduced order models for basis functions that are custom-made to represent the solution of partial differential equations. Partial differential equations (PDEs) are one of the main mathematical tools for describing physical phenomena. However, due to either efficiency or necessity, for many real-world problems, we are interested in constructing reduced order models (ROMs) which focus only on the explicit computation of subsets of the active spatio-temporal scales in the problem, while treating the interaction with the rest of the scales approximately. The task of accurate representation of such interactions (usually called memory terms) constitutes a vast area of research known as model reduction. PI Stinis has significant expertise in the construction of ROMs for complex systems. In addition, in recent work with the project key participant Qadeer, they have utilized machine learning to acquire custom-made basis functions (CBFs) to expand the solutions of PDEs. In the proposed work, we will merge the two concepts by constructing ROMs for subsets of the CBFs needed to represent the solution of a PDE. Specifically, we will use the Mori-Zwanzig model reduction formalism to construct ROMs for subsets of CBFs for nonlinear PDEs of various complexity, as well as investigate the usage of CBFs in the spectral vanishing viscosity method for problems that can form shocks in finite time. The outcome of the research is aimed to be proof-of-concept about a novel synergistic approach between model reduction and machine learning, thus advancing the field of scientific machine learning. Such a capability will benefit the efficient modeling of physical systems appearing in various areas of interest to the DOE.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Dynamic density functional theory of polymers with salt in electric fields

Here we present a dynamic density functional theory for modeling the effects of applied electric fields on the local structure of polymers with added salt (polymer electrolytes). Time-dependent equations for the local electrostatic potential and volume fractions of polymer, cation, and anion of added salt are developed using the principles of linear irreversible thermodynamics. For such a development, a field theoretic description of the free energy of polymer melts doped with salts is used, which captures the effects of local variations in the dielectric function. Connections of the dynamic density functional theory with experiments are established by relating the three phenomenological Onsager’s transport coefficients of the theory to the mutual diffusion of electrolyte, ionic conductivity, and transference number of one of the ions. The theory is connected with a statistical mechanical model developed by Bearman and Kirkwood [J. Chem. Phys. 28, 136 (1958)] after relating the three transport coefficients to friction coefficients. The steady-state limit of the dynamic density functional theory is used to understand the effects of dielectric inhomogeneity on the phase separation in polymer electrolytes. The theory developed here provides not only a way to connect with experiments but also to develop multi-scale models for studying connections between local structure and ion transport in polymer electrolytes.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2024

Spike-and-Slab Shrinkage Priors for Structurally Sparse Bayesian Neural Networks

Network complexity and computational efficiency have become increasingly significant aspects of deep learning. Sparse deep learning addresses these challenges by recovering a sparse representation of the underlying target function by reducing heavily overparameterized deep neural networks. Specifically, deep neural architectures compressed via structured sparsity (e.g., node sparsity) provide low-latency inference, higher data throughput, and reduced energy consumption. In this article, we explore two well-established shrinkage techniques, Lasso and Horseshoe, for model compression in Bayesian neural networks (BNNs). To this end, we propose structurally sparse BNNs, which systematically prune excessive nodes with the following: 1) spike-and-slab group Lasso (SS-GL) and 2) SS group Horseshoe (SS-GHS) priors, and develop computationally tractable variational inference, including continuous relaxation of Bernoulli variables. We establish the contraction rates of the variational posterior of our proposed models as a function of the network topology, layerwise node cardinalities, and bounds on the network weights. Furthermore, we empirically demonstrate the competitive performance of our models compared with the baseline models in prediction accuracy, model compression, and inference latency.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Comparison and multi-model inference of excess risks models for radiation-related solid cancer

In assessments of detrimental health risks from exposures to ionising radiation, many forms of risk to dose–response models are available in the literature. The usual practice is to base risk assessment on one specific model and ignore model uncertainty. The analysis illustrated here considers model uncertainty for the outcome all solid cancer incidence, when modelled as a function of colon organ dose, using the most recent publicly available data from the Life Span Study on atomic bomb survivors of Japan. Seven recent publications reporting all solid cancer risk models currently deemed plausible by the scientific community have been included in a model averaging procedure so that the main conclusions do not depend on just one type of model. The models have been estimated with different baselines and presented for males and females at various attained ages and ages at exposure, to obtain specially computed model-averaged Excess Relative Risks (ERR) and Excess Absolute Risks (EAR). Monte Carlo simulated estimation of uncertainty on excess risks was accounted for by applying realisations including correlations in the risk model parameters. Three models were found to weight the model-averaged risks most strongly depending on the baseline and information criteria used for the weighting. Fitting all excess risk models with the same baseline, one model dominates for both information criteria considered in this study. Based on the analysis presented here, it is generally recommended to take model uncertainty into account in future risk analyses.

63 RADIATION, THERMAL, AND OTHER ENVIRON. POLLUTAN↗
DOE OSTI2022

Engineering infinite-range SU($\textit{n}$) interactions with spin-orbit-coupled fermions in an optical lattice

Here, we study multilevel fermions in an optical lattice described by the Hubbard model with on-site SU($\textit{n}$) -symmetric interactions. We show that in an appropriate parameter regime this system can be mapped onto a spin model with all-to-all SU($\textit{n}$)-symmetric couplings. Raman pulses that address internal spin states modify the atomic dispersion relation and induce spin-orbit coupling, which can act as a synthetic inhomogeneous magnetic field that competes with the SU($\textit{n}$) exchange interactions. We investigate the mean-field dynamical phase diagram of the resulting model as a function of $\textit{n}$ and different initial configurations that are accessible with Raman pulses. Consistent with previous studies for $\textit{n}$ = 2 , we find that for some initial states the spin model exhibits two distinct dynamical phases that obey simple scaling relations with $\textit{n}$. Moreover, for $\textit{n}$ > 2 we find that dynamical behavior can be highly sensitive to initial intraspin coherences. Our predictions are readily testable in current experiments with ultracold alkaline-earth-metal(-like) atoms.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2024

Data-Driven Supervised Dimension Reduction for Scientific Discovery (LDRD QTI Report)

This report summarizes the findings of a four months FY24 Advanced Science & Technology (AS&T) LDRD Quick Targeted Investigation (QTI) project focused on the exploration of supervised dimension reduction approaches based on autoencoders. Autoencoders have been extensively employed in literature for unsupervised learning tasks, however, their use for supervised regression tasks, which are common within scientific applications, has been limited. Motivated by linear dimension reduction strategies like Active Subspaces and Adaptive Basis, we explored the possibility of employing autoencoders to discover a non-linear manifold able to represent the original function in fewer dimensions. In this report, we discuss a neural network architecture and we perform a numerical campaign on several problems ranging from simple two-dimensional functions to a model problem for magnetohydrodynamics in five dimensions. In our preliminary results, we show that the proposed approach is found to be superior to linear dimension reduction strategies in representing the target function even with a single latent variable.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Application of functional genomics for domestication of novel non-model microbes

Abstract With the expansion of domesticated microbes producing biomaterials and chemicals to support a growing circular bioeconomy, the variety of waste and sustainable substrates that can support microbial growth and production will also continue to expand. The diversity of these microbes also requires a range of compatible genetic tools to engineer improved robustness and economic viability. As we still do not fully understand the function of many genes in even highly studied model microbes, engineering improved microbial performance requires introducing genome-scale genetic modifications followed by screening or selecting mutants that enhance growth under prohibitive conditions encountered during production. These approaches include adaptive laboratory evolution, random or directed mutagenesis, transposon-mediated gene disruption, or CRISPR interference (CRISPRi). Although any of these approaches may be applicable for identifying engineering targets, here we focus on using CRISPRi to reduce the time required to engineer more robust microbes for industrial applications. One-Sentence Summary The development of genome scale CRISPR-based libraries in new microbes enables discovery of genetic factors linked to desired traits for engineering more robust microbial systems.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2021

Advances in constraining intrinsic alignment models with hydrodynamic simulations

We use galaxies from the illustristng, massiveblack-ii, and illustris-1 hydrodynamic simulations to investigate the behaviour of large scale galaxy intrinsic alignments. Our analysis spans four redshift slices over the approximate range of contemporary lensing surveys z = 0–1. We construct comparable weighted samples from the three simulations, which we then analyse using an alignment model that includes both linear and quadratic alignment contributions. Our data vector includes galaxy–galaxy, galaxy–shape, and shape–shape projected correlations, with the joint covariance matrix estimated analytically. In all of the simulations, we report non-zero IAs at the level of several σ. For a fixed lower mass threshold, we find a relatively strong redshift dependence in all three simulations, with the linear IA amplitude increasing by a factor of ~2 between redshifts z = 0 and z = 1. We report no significant evidence for non-zero values of the tidal torquing amplitude, A 2 , in TNG, above statistical uncertainties, although MBII favours a moderately negative A 2 ~ –2. Examining the properties of the TATT model as a function of colour, luminosity and galaxy type (satellite or central), our findings are consistent with the most recent measurements on real data. We also outline a novel method for constraining the TATT model parameters directly from the pixelized tidal field, alongside a proof-of-concept exercise using TNG. This technique is shown to be promising, although comparison with previous results obtained via other methods is non-trivial.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2022

Anisotropic angular scattering models of elastic electron-neutral collisions for Monte Carlo plasma simulations

Many laboratory and industrial plasma applications require accurate modeling techniques to understand the interplay between microscopic and macroscopic processes. A prime example of this interplay is how particle and Monte Carlo (MC) simulation codes describe angular scattering of electrons following elastic scattering events. The forward peaked nature of high energy electron elastic scattering is relatively trivial to accurately describe in plasma simulations. However, for lower energy collisions, which produce near isotropic or backward peaked differential cross sections, there is not a strong consensus among the plasma modeling community on how to best describe these angular scattering trends. Here, in this study, we propose a systematic method to approximate the aforementioned non-trivial angular scattering behavior with a formula that can be readily implemented in particle-in-cell (PIC) and/or MC plasma simulation codes. The present approach is specifically applied to fusion relevant atomic hydrogen and helium, as well as for molecular hydrogen, and results are also applicable to the atomic isotopes and homonuclear molecular isotopologues of these species. Comparisons between the present angular distribution function and benchmark scattering data were used to validate the proposed models. In addition, two-term Boltzmann calculations and PIC direct simulation MC simulations revealed that the proposed angular distribution function is accurate, agreeing very well with benchmark convergent close-coupling scattering calculations, and electron transport measurements. These studies confirmed that the present angular distribution function model can be utilized without the need of renormalization to the momentum transfer cross section (as opposed to using the elastic scattering integrated cross section), which has been suggested by several studies in order to correct for deficient angular scattering models, and to agree with transport measurements. Hence, the present anisotropic angular scattering model can be utilized to accurately model the momentum transfer as well as the electron trajectories of elastic collisions.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2022

The eROSITA Final Equatorial-Depth Survey (eFEDS): The AGN catalog and its X-ray spectral properties

The eROSITA Final Equatorial Depth Survey (eFEDS), observed with eROSITA ahead of its planned 4-yr all-sky survey, is the largest contiguous-field X-ray survey at present. It yielded a large sample of X-ray sources with very rich multiband photometric and spectroscopic coverage. We present here the eFEDS active galactic nuclei (AGN) catalog and the eROSITA X-ray spectral properties of the eFEDS sources. Using a Bayesian method, we performed a systematic X-ray spectral analysis for all the eFEDS sources. We adopted multiple spectral models, including single-component power-law or hot-plasma models and double-component models of a power law plus soft excess. We investigated the capacity of eROSITA X-ray spectra for constraining AGN spectral shapes through a detailed analysis of the posterior parameter probability distribution functions. Hierarchical Bayesian modeling was used to recover the spectral parameter distribution of the sample. The source fluxes and luminosities were measured from the posterior of the spectral fitting. The eFEDS AGN catalog (22 079 sources) comprises ~80% of the eFEDS point sources. Despite a large number of faint sources, our spectral fitting provides reasonable measurements of spectral shapes and intrinsic luminosities for a majority of the sources. Because of sample selection bias, this AGN catalog is dominated by X-ray unobscured sources, with an obscured (logN H > 21.5) fraction of 8%; the power-law emission of the hot corona is also relatively soft, with a typical slope of 2.0. For type-I AGN, the X-ray emission is well correlated with the UV emission with the usual anticorrelation between the X-ray to UV spectral slope α OX and the UV luminosity. The X-ray spectral properties measured with various models are presented for all the eFEDS sources.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2021

Model-based Cyber-attack Detection for Voltage Source Converters in Island Microgrids

With the upgrading of communication and control in microgrids, cyber threats on the power converters are increasing. Here, in this paper, a novel model-based cyber-attack detection methodology is proposed for each voltage source converter in microgrids. The Harmonics State Space Matrix (H-Matrix) is used to build the closed-loop transfer function, providing model-based estimation with the grid voltage and control reference. Then, the space phase model (SPM) is used to calculate residual to detect cyber-attacks in voltage source converter (VSC). As controller and converter parameters are included in the H-matrix, the proposed method also could detect cyber-attacks when voltage fluctuation occurs in the grid. To verify the feasibility of the proposed method, several attacks targeting VSC in an islanded microgrid are simulated in MATLAB. Simulation and detection results are introduced to demonstrate the resilience and feasibility of the proposed method.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2022

Accelerating Multiscale Materials Modeling with Machine Learning

The focus of this project is to accelerate and transform the workflow of multiscale materials modeling by developing an integrated toolchain seamlessly combining DFT, SNAP, LAMMPS, (shown in Figure 1-1) and a machine-learning (ML) model that will more efficiently extract information from a smaller set of first-principles calculations. Our ML model enables us to accelerate first-principles data generation by interpolating existing high fidelity data, and extend the simulation scale by extrapolating high fidelity data (10 2 atoms) to the mesoscale (10 4 atoms). It encodes the underlying physics of atomic interactions on the microscopic scale by adapting a variety of ML techniques such as deep neural networks (DNNs), and graph neural networks (GNNs). We developed a new surrogate model for density functional theory using deep neural networks. The developed ML surrogate is demonstrated in a workflow to generate accurate band energies, total energies, and density of the 298K and 933K Aluminum systems. Furthermore, the models can be used to predict the quantities of interest for systems with more number of atoms than the training data set. We have demonstrated that the ML model can be used to compute the quantities of interest for systems with 100,000 Al atoms. When compared with 2000 Al system the new surrogate model is as accurate as DFT, but three orders of magnitude faster. We also explored optimal experimental design techniques to choose the training data and novel Graph Neural Networks to train on smaller data sets. These are promising methods that need to be explored in the future.

36 MATERIALS SCIENCE↗
DOE OSTI2026

Co-Design of Charge Transport Superhighways to Connect Catalytic Sites in Soft Photoelectrochemical Systems

Efficient photon-to-electron-to-molecule conversion requires multi-length scale control over charge transport pathways, where electronic charges are delivered to catalytic sites under high mass transport flux. A fundamental question is how can we co-design charge transport pathways to promote efficient charge transfer to/from catalytic sites in complex three-dimensional architectures? Soft conducting polymer systems offer exceptional promise to provide three-dimensional charge transport networks, where electrolyte (ion and solvent) can interdiffuse to promote long-lived charge carriers and the molecular nature allows for strategic synthetic design of catalytic sites. Herein we combine theoretical and experimental approaches to investigate the earliest stages of photoelectrochemical deposition of near-surface catalytic sites (Pt) on soft bulk heterojunction polymeric semiconductors composed of a prototype donor (PTB7-Th) and a prototype acceptor (N2200) as a model system towards better understanding molecular catalyst-polymer site interactions. We focus initially on photoelectrochemical deposition of low Pt loadings, nanoparticle sizes (formed by progressive nucleation) below 20 nm, for both density functional theory (DFT) modeling studies and for spectroscopic characterization using surface-sensitive X-ray and UV-photoemission (XPS/UPS). DFT modeling of “n-type” N2200 slabs reveal for the first time that sulfur atoms in the thiophene units serve as the lowest-energy adsorption sites for single Pt atoms, while larger Pt clusters engage more complexly with both thiophene and naphthalene diimide (NDI) core sites. Changes in chemical composition observed by X-ray photoelectron spectroscopy (XPS) support the DFT predictions, and the angle-resolved measurements reveal that Pt nucleation initiates at subsurface sites which appear to be localized active domains that promote charge transport/transfer and enable vertical growth toward the surface. These results suggest that light-activated Pt nanoparticle deposition decorates energetically distinct sites, where photoactivity is dictated by the local energetics of those sites, and the fact that they represent the termini of charge transport “super-highways” – a small percentage of the total volume of the donor/acceptor polymeric active layer which carries most of the photocurrent generated during both Pt deposition and photoelectrochemical HER. We posit that these initial studies provide a foundational strategy for design of catalytic sites in the near surface regions of complex polymeric materials and advancing soft semiconductor-based photoelectrochemical systems. Achieving a nanometer-scale understanding of catalyst deposition and the impact of local composition and energetics on that placement, should ultimately provide the design guidelines (co-design) for a broad array of catalysts at sites that optimize that efficiency and maximize platform durability.

14 SOLAR ENERGY↗
DOE OSTI2023

Deep learning closure models for large-eddy simulation of flows around bluff bodies

Near-wall flow simulation remains a central challenge in aerodynamics modelling: Reynolds-averaged Navier–Stokes predictions of separated flows are often inaccurate, and large-eddy simulation (LES) can require prohibitively small near-wall mesh sizes. A deep learning (DL) closure model for LES is developed by introducing untrained neural networks into the governing equations and training in situ for incompressible flows around rectangular prisms at moderate Reynolds numbers. The DL-LES models are trained using adjoint partial differential equation (PDE) optimization methods to match, as closely as possible, direct numerical simulation (DNS) data. They are then evaluated out-of-sample – for aspect ratios, Reynolds numbers and bluff-body geometries not included in the training data – and compared with standard LES models. The DL-LES models outperform these models and are able to achieve accurate LES predictions on a relatively coarse mesh (downsampled from the DNS mesh by factors of four or eight in each Cartesian direction). We study the accuracy of the DL-LES model for predicting the drag coefficient, near-wall and far-field mean flow, and resolved Reynolds stress. A crucial challenge is that the LES quantities of interest are the steady-state flow statistics; for example, a time-averaged velocity component $\langle {u}_i\rangle (x) = \lim _{t \rightarrow \infty } ({1}/{t}) \int _0^t u_i(s,x)\, {\rm d}s$ . Calculating the steady-state flow statistics therefore requires simulating the DL-LES equations over a large number of flow times through the domain. It is a non-trivial question whether an unsteady PDE model with a functional form defined by a deep neural network can remain stable and accurate on $t \in [0, \infty )$ , especially when trained over comparatively short time intervals. Our results demonstrate that the DL-LES models are accurate and stable over long time horizons, which enables the estimation of the steady-state mean velocity, fluctuations and drag coefficient of turbulent flows around bluff bodies relevant to aerodynamics applications.

Mechanics↗